Explainable artificial intelligence models using SHAP enhanced CatBoost, Bi-GRU with attention, and Tab Transformer

Abstract

Abstract Insurance fraud detection remains challenging to predict in reality because claims data is often uneven among classes, and the information concerning claims is often multidimensional and nonhomogeneous. The present research used a unified evaluation framework to assess the predictive and interpretive capabilities of three distinct model families: CatBoost (tree-based ensemble learning), Bi-GRU with Attention (sequence-oriented learning), and TabTransformer (categorical feature contextual). The model families were tested using a standardised experimental protocol.The study is novel in the sense of a cross-model interpretability framework that unites Shapley Additive Explanation (SHAP)-based feature attribution with attention-based contextual analysis to enable a clear comparison of model reasoning between the suggested frameworks. The data on which the experiments were done consisted of 4,000 life insurance claims that were characterized in terms of 83 attributes. Common preprocessing procedures like missing values, scaling numerical variables, and selecting highly correlated variables were used before training the models. Experimentally, CatBoost is proven to be the most precise on legitimate claims, Bi-GRU is the most recall on fraudulent claims, and TabTransformer is the best in terms of tradeoff between accuracy, interpretability, and computational efficiency. Practical characteristics such as the quantity of claim, tenure in a policy, and diagnosis were repeatedly emphasized in both SHAP and attention analyses. Combined, the current research study provides a consistent and explainable benchmark that may be applied to conduct fraud detection research reliably and assist practitioners in choosing models that are accurate and understandable.